Building more with GPT-5.1-Codex-Max

Building more with GPT-5.1-Codex-Max | OpenAI

November 19, 2025

Building more with GPT‑5.1‑Codex‑Max

Get started

$ npm i -g @openai/codex

Loading…

Share

Introduction

We’re introducing GPT‑5.1‑Codex‑Max, our new frontier agentic coding model, available in Codex today. GPT‑5.1‑Codex‑Max is built on an update to our foundational reasoning model, which is trained on agentic tasks across software engineering, math, research, and more. GPT‑5.1‑Codex‑Max is faster, more intelligent, and more token-efficient at every stage of the development cycle–and a new step towards becoming a reliable coding partner.

GPT‑5.1‑Codex‑Max is built for long-running, detailed work. It’s our first model natively trained to operate across multiple context windows through a process called compaction, coherently working over millions of tokens in a single task. This unlocks project-scale refactors, deep debugging sessions, and multi-hour agent loops.

GPT‑5.1‑Codex‑Max is available in Codex today for use in the CLI, IDE extension, cloud, and code review, and API access is coming soon.

Frontier coding capabilities

GPT‑5.1‑Codex‑Max was trained on real-world software engineering tasks, like PR creation, code review, frontend coding, and Q&A and outperforms our previous models on many frontier coding evaluations. The model’s gains on benchmarks also come with improvements to real-world usage: GPT‑5.1‑Codex‑Max is the first model we have trained to operate in Windows environments, and the model’s training now includes tasks designed to make it a better collaborator in the Codex CLI.

• All evals were run with compaction enabled at Extra High reasoning effort* Terminal-Bench2.0 ran with Codex CLI in the Laude Institute Harbor harness⁠

Speed and cost

GPT‑5.1‑Codex‑Max shows significant improvements in token efficiency due to more effective reasoning. On SWE-bench Verified, GPT‑5.1‑Codex‑Max with ‘medium’ reasoning effort achieves better performance than GPT‑5.1‑Codex with the same reasoning effort, while using 30% fewer thinking tokens. For non-latency-sensitive tasks, we’re also introducing a new Extra High (‘xhigh’) reasoning effort, which thinks for an even longer period of time for a better answer. We still recommend medium as the daily driver for most tasks.

We expect the token efficiency improvements to translate to real-world savings for developers.

For example, GPT‑5.1‑Codex‑Max is able to produce high quality frontend designs with similar functionality and aesthetics, but at much lower cost than GPT‑5.1‑Codex.

Prompt: Generate a single self-contained browser app that renders an interactive CartPole RL sandbox with canvas graphics, a tiny policy-gradient controller, metrics, and an SVG network visualizer.

Features

• Must be able to actually train a policy to make model better at cart pole
• Visualizer for the activations/weights when the model is training or at inference
• Steps in the episode, rewards this episode
• Last survival time and best survival time in steps

Save to index.html

Long-running tasks

Compaction enables GPT‑5.1‑Codex‑Max to complete tasks that would have previously failed due to context-window limits, such as complex refactors and long-running agent loops by pruning its history while preserving the most important context over long horizons. In Codex applications, GPT‑5.1‑Codex‑Max automatically compacts its session when it approaches its context window limit, giving it a fresh context window. It repeats this process until the task is completed.

The ability to sustain coherent work over long horizons is a foundational capability on the path toward more general, reliable AI systems. GPT‑5.1‑Codex‑Max can work independently for hours at a time. In our internal evaluations, we’ve observed GPT‑5.1‑Codex‑Max work on tasks for more than 24 hours. It will persistently iterate on its implementation, fix test failures, and ultimately deliver a successful result.

In this example, GPT‑5.1‑Codex‑Max is independently refactoring the Codex CLI open source repository.

As the session length approaches the model’s context-window, it automatically compacts the session to free up space to continue the task without losing progress.

The video has been trimmed and sped up for clarity.

Building safe and trustworthy AI agents

GPT‑5.1‑Codex‑Max performs significantly better on evaluations that require sustained, long-horizon reasoning. Because it can coherently work across multiple context windows using compaction, the model delivers improved results on challenges in areas like long-horizon coding and cybersecurity. We analyzed the results of this model’s performance on first- and third-party evaluations in the GPT‑5.1‑Codex‑Max system card⁠.

GPT‑5.1‑Codex‑Max does not reach High capability on Cybersecurity under our Preparedness Framework ⁠ but it is the most capable cybersecurity model we’ve deployed to date and agentic cybersecurity capabilities are rapidly evolving. As a result, we are taking steps to prepare for High capability on Cybersecurity and are enhancing our safeguards in the cyber domain and working to ensure that defenders can benefit from these improved capabilities through programs like Aardvark⁠.

When we launched GPT‑5‑Codex, we implemented dedicated cybersecurity-specific monitoring to detect and disrupt malicious activity. While we have not observed a meaningful increase in scaled abuse, we are preparing additional mitigations for advanced capabilities. Our teams have already disrupted cyber operations⁠ attempting to misuse our models, and suspicious activity is routed for review through our policy monitoring systems.

Codex is designed to run in a secure sandbox by default: file writes are limited to its workspace, and network access is disabled unless a developer turns it on. We recommend keeping Codex in this restricted-access mode, since enabling internet or web search can introduce prompt-injection⁠ risks from untrusted content.

As Codex becomes more capable of long-running tasks, it is increasingly important for developers to review the agent’s work before making changes or deploying to production. To assist with this, Codex produces terminal logs and cites its tool calls and test results. While its code reviews reduce the risk of deploying model or human produced bugs to production, Codex should be treated as an additional reviewer and not a replacement for human reviews.

Cybersecurity capabilities can be used for both…